---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
language:
  - en
tags:
  - math
  - reasoning
  - reasoning-compression
  - self-distillation
  - crisp
base_model:
  - Qwen/Qwen3-8B
datasets:
  - pb09204048/CRISP
---

# CRISP-Qwen3-8B-v2

**Qwen3-8B** trained with **CRISP** (Compressed Reasoning via Iterative Self-Policy Distillation)
using the **v2** conciseness teacher. Step-99 checkpoint.

**Paper:** https://arxiv.org/abs/2603.05433

CRISP teaches a reasoning model to think concisely by distilling its own concise behavior back into
itself: the teacher is the *same* model conditioned on a conciseness instruction, the student has no
instruction, and training minimizes per-token reverse KL from student to teacher on the student's own
rollouts (teacher refreshed every `M=50` steps). No ground-truth answers, token budgets, or
difficulty estimators enter the loss.

This checkpoint uses the **v2** teacher prompt:
**v2 (difficulty-aware, default):** adds a caveat to *not over-compress* hard/multi-step problems (keep case analysis, edge cases, a final check).

Other CRISP checkpoints: [Qwen3-8B](https://huggingface.co/pb09204048/CRISP-Qwen3-8B-v1) ([v2](https://huggingface.co/pb09204048/CRISP-Qwen3-8B-v2)),
[Qwen3-14B](https://huggingface.co/pb09204048/CRISP-Qwen3-14B-v1) ([v2](https://huggingface.co/pb09204048/CRISP-Qwen3-14B-v2)),
[DeepSeek-R1-Distill-Llama-8B](https://huggingface.co/pb09204048/CRISP-DeepSeek-R1-Distill-Llama-8B-v1) ([v2](https://huggingface.co/pb09204048/CRISP-DeepSeek-R1-Distill-Llama-8B-v2)).
Training data: [pb09204048/CRISP](https://huggingface.co/datasets/pb09204048/CRISP).

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("pb09204048/CRISP-Qwen3-8B-v2")
model = AutoModelForCausalLM.from_pretrained("pb09204048/CRISP-Qwen3-8B-v2", device_map="auto")
```

## Benchmark results (Qwen3-8B)

Accuracy (mean@8, %) and token reduction (Red., % vs. base) at a 30K-token budget. Math is scored with
a dual-path grader (`Answer:` or `\boxed{}`); GPQA-Diamond and MMLU use exact letter-match. This
model is the **CRISP (v2)** row.

| Setting | MATH-500 | AIME 2024 | AIME 2025 | GPQA-D | MMLU |
|---------|----------|-----------|-----------|--------|------|
| Base | 95.7 / — | 76.2 / — | 70.4 / — | 61.5 / — | 81.9 / — |
| Concise prompt (v2) | 94.2 / 21.5% | 74.6 / 9.8% | 63.7 / 5.3% | 59.5 / 30.7% | 82.8 / 26.8% |
| Concise prompt (v1) | 95.6 / 38.9% | 74.2 / 20.2% | 62.1 / 13.9% | 56.8 / 29.5% | 83.0 / 26.8% |
| **CRISP (v2)** | 95.7 / 31.6% | 75.0 / 17.1% | 65.8 / 17.5% | 58.3 / 17.2% | 81.2 / 22.4% |
| **CRISP (v1)** | 95.7 / 56.9% | 72.9 / 32.9% | 58.8 / 28.4% | 58.5 / 36.2% | 80.9 / 44.7% |

## Citation

```bibtex
@article{sang2026crisp,
  title={Crisp: Compressed reasoning via iterative self-policy distillation},
  author={Sang, Hejian and Xu, Yuanda and Zhou, Zhengze and He, Ran and Wang, Zhipeng and Sun, Jiachen},
  journal={arXiv preprint arXiv:2603.05433},
  year={2026}
}
```